PES Benchmark v0.2 | MyShape Protocol | July 2026
There's a quiet assumption powering the AI revolution: that anything a human can generate, a model can learn to copy. Faces. Voices. Writing style. Fingerprints. Every static biometric has fallen, one by one, to diffusion models and generative adversarial networks.
But motion is different.
Motion isn't a snapshot. It's a continuous, real-time biological process — muscles contracting, neurons firing, stretch reflexes oscillating at 8–12 Hz. It has physics underneath it, and physics leaves a trace that pure statistics cannot replicate.
That's the hypothesis we set out to test. And the results surprised even us.
We built an engine called PES (Presence Entropy Scoring). It takes raw webcam footage, runs it through MediaPipe Pose to extract 18 3D skeleton landmarks, then scores the motion on a 0–1 "biological entropy" scale.
The question was simple: can entropy alone tell a real human from AI-generated movement?
We collected 231 samples: 81 from real humans performing five different action types (waving, nodding, walking, gesturing, standing still), and 150 from three categories of AI-generated motion — random walk, spline interpolation, and near-static models.
Every sample went through the same pipeline. No special treatment. No cherry-picking.
The engine never once mistook AI for human.
Human | AI (3 models) | |
|---|---|---|
Mean PES | 0.59 | 0.00–0.31 |
Pass rate (>0.5) | 88.9% | 0.0% |
Precision: 100% (95% confidence interval [97.6%, 100%]) Cohen's d: 10.4
For those unfamiliar with Cohen's d: in psychology, d > 0.8 is considered a "large effect." At d = 10.4, the distributions of human and AI motion don't just differ — they barely overlap.
(在此处插入你的图片：image_37ad1f.jpg)
The 9 humans that didn't pass were standing completely still. Minimal motion produces minimal entropy. That's not a failure — it's exactly what the model predicts. The system isn't measuring whether you look human. It's measuring whether you move like one.
What makes this different from black-box ML is that each dimension has a falsifiable biological hypothesis behind it. We're not training a classifier to "detect deepfakes." We're testing whether specific properties of the human nervous system are measurable through consumer hardware.
Real camera frames arrive with 28–38 milliseconds of jitter between captures. Your webcam isn't a metronome — USB bus contention, OS scheduling, and sensor readout all inject variance. AI-generated video? Exactly 33.33ms per frame. Perfect clock. We measure the coefficient of variation in inter-frame intervals. Humans vary. Models don't.
Your muscles aren't perfectly still, even when you try to be. The stretch reflex loop produces micro-oscillations at 8–12 Hz with amplitudes of roughly 0.01–0.02 pixels. AI motion, trained with L2 loss, smooths away exactly this signal. It doesn't know that deviation IS the signal. The better the model gets at generating visually convincing motion, the more aggressively it smooths — and the more detectable it becomes. AI can't do both.
When you raise your arm, your shoulder and wrist accelerate together. Your motor cortex coordinates muscle groups as coupled oscillators. AI models generate each joint independently. There's no "nervous system" connecting them. The cross-correlation matrix of joint jerk is nearly zero.
We discovered a pipeline artifact: AI videos replayed on a screen and re-recorded through the same webcam scored higher than real humans — 0.65 vs 0.59 mean PES.
It's a false positive factory. We documented this in Section 4 of the report without sugarcoating. Science that only reports successes isn't science. The next experiment is designed specifically to disambiguate: is this a fundamental detection failure or a pipeline artifact we can filter out with challenge-response protocols?
We are not claiming that PES proves humanness. We are reporting a single, narrow, falsifiable finding: biologically-derived motion entropy appears to be a measurable, testable signal that distinguishes real human movement from current AI-generated alternatives.
If this finding replicates at scale, it suggests that the physical constraints of biological bodies — muscle tension, neural coupling, reflex arcs — produce entropy signatures that statistical models, by their mathematical nature, cannot reproduce.
Larger human sample — 300+ subjects.
Real AI video — Testing Kling, Runway, NVIDIA Cosmos through the AGNES API.
Challenge-response — Implementing interactive verification.
Open-source the benchmark — The engine and test suite are MIT licensed.
The engine is open source. 172 tests. TypeScript + Rust.
👉 github.com/myshapeprotocol/myshape-protocol
MyShape Protocol is building sovereign 3D identity infrastructure for the decentralized human. We research motion-signature verification, zero-knowledge presence proofs, and the entropy gap between biological and synthetic motion.
Contact: protocol@myshape.com